We didn’t just hunt alpha; we rewired the game. That’s what Moore Threads co-founder Wang Dong did last week when he declared there’s no universal chip for inference, only a combination of solutions. For a moment, I thought he was talking about crypto. His words hit me like a flashback to 2017, sitting in a Jakarta coffee shop, auditing a Solidity contract that promised to be “the one chain to rule them all.” It wasn’t. And the pattern repeats.
Hook – A single statement, a whole industry shaken. Wang Dong, a veteran from China’s GPU wars, told the audience at a tech summit: “The inference market does not have a universal chip; it has a combination of solutions.” He flipped the script on the monolithic GPU narrative. For the crypto faithful, this sounds familiar. How many times did we hear “Bitcoin is the only chain” or “Ethereum will solve everything”? Both failed to scale for all use cases. The modular thesis – rollups, L2s, specialized data availability layers – won the debate. Now the same fragmentation is hitting AI hardware. And I see a deeper parallel: the way we design trust in decentralized systems mirrors how we design efficiency in inference machines.
Context – The protocol background. Wang Dong’s company, Moore Threads, builds GPUs for AI workloads in a market dominated by NVIDIA. His argument is deceptively simple: each inference scenario – low-latency chat, high-throughput batch, streaming code completion – demands a different silicon sweet spot. A single chip can’t optimize for all. So he champions a future of “Inference Service Providers” (ISPs) that mix and match GPUs, ASICs, and even NPUs to deliver the best cost-performance for each model. This is the very same logic that drove me from auditing smart contracts in 2017 to building BlockJakarta in 2024: trust is not a monolith. It’s a stack. From core dev trenches to community heartbeat, I’ve watched blockchains evolve from monolithic L1s to interconnected L2s. Wang Dong’s ISP vision is the crypto modular thesis, but for compute.
Core – The technical and values analysis. Let’s break it down. In blockchain, we have the modular stack: L1 for settlement, L2 for execution, DA layer for data. Each layer is a specialized chip for a specific function. No single chain can handle infinite throughput, finality, and privacy simultaneously – just as no single chip can handle ultra-low latency inference and massive parallel processing at the same time. Wang Dong’s “combination of solutions” mirrors exactly the rollup-centric roadmap. During DeFi Summer in 2020, I launched a localized AMM called UniBarter in Jakarta. I thought one protocol could serve all Indonesian traders. It failed. The community splintered into yield farmers on Ethereum and arbitrageurs on Binance Smart Chain. Fragmentation wasn’t a bug; it was the market expressing its need for specialized trust primitives. Likewise, inference workloads are splintering. Large language models like GPT-4 need big GPU clusters for batch inference. Small models like a code assistant need low latency from fast SRAM-heavy chips (like Groq’s LPU). And models quantized to 3 bits might run best on custom ASICs. Wang Dong isn’t just making a technical argument; he’s making a values argument. He is saying: stop chasing a single perfect chip. Instead, embrace a heterogeneous ecosystem where each component does one thing well, and the trust (in this case, performance) emerges from the combination. This is the same philosophical shift that moved DeFi from “one DEX to rule them all” to a world of Uniswap, Curve, Balancer, each optimized for a specific liquidity profile. I’ve seen it happen in crypto. Now I see it happening in hardware.
Let’s go deeper. During the Terra/Luna collapse in 2022, I retreated to my apartment and wrote a 50-page analysis of why trustless systems based on infinite growth fail. That experience taught me that every monolithic trust system has a single point of failure – usually in its assumption of universal applicability. The same applies to inference chips. NVIDIA’s H100 is an incredible universal GPU, but it’s overkill for a simple chatbot inference task. Wang Dong’s argument is that relying on one chip is like relying on one blockchain for all transactions: it works until the demand pattern shifts. In crypto, we solved this with modular chains that share a security layer but optimize execution separately. In inference, we are starting to see the same: different chips for different tasks, all managed by a software orchestration layer (the inference server equivalent of a rollup sequencer).
But here is where my grounded skepticism kicks in. “Combination of solutions” sounds elegant until you try to run it. I’ve built and managed multi-chain liquidity bridges. The complexity is enormous. The same will haunt ISPs. Wang Dong’s vision requires a unified software stack that can dynamically allocate a query to the cheapest, fastest available chip while maintaining output consistency. That’s harder than it looks. In crypto, moving between L2s often incurs delays, fees, and security assumptions. In inference, moving between different GPU architectures can cause numerical differences due to varying precision (FP16 vs BF16) and operator implementations. A model that outputs 0.1 on NVIDIA might output 0.12 on Moore Threads. For a creative writing assistant, that’s fine. For a medical diagnosis model, it’s catastrophic. This is the “model fragmentation” problem – analogous to the fragmentation of state across rollups. Wang Dong didn’t address this. And that’s a risk.
Contrarian – The counter-intuitive angle. Here is the blind spot: Wang Dong’s combination thesis actually strengthens NVIDIA’s position, not weakens it. If the market demands seamless heterogeneity, the winning solution is the one that offers the best orchestration layer – and NVIDIA already has it with CUDA, TensorRT-LLM, and its AI Enterprise stack. They just need to open it up a bit. Meanwhile, Moore Threads and other Chinese GPU makers are years behind in software maturity. I’ve audited enough smart contracts to know: a new chain with great L1 speed but poor EVM compatibility will fail. Similarly, a new chip with great specs but poor PyTorch integration will struggle. My experience in the DeFi trenches taught me that developer experience and trust in the toolchain dominate network effects. The same applies here. Wang Dong is essentially saying “let’s build a multi-chain ecosystem of chips.” But the operator (ISP) will face significant challenges in security, data isolation, and performance consistency. The Terra collapse taught me that trustless systems need robust fallback mechanisms. In inference, if the combination fails due to a bug in the orchestrator, the entire service goes down. The contrarian truth: combination solutions increase surface area for failure. Modular blockchain architecture introduced new attack vectors on bridges and sequencers. Modular inference architecture will introduce new attack vectors on the scheduling layer and hardware-specific bugs. I expect the first major ISP outage to happen within 12 months, traced to a subtle incompatibility between two chips from different vendors.
Takeaway – So where does this leave us? Wang Dong is right about the trend: the future is not a single universal chip, just as the future of crypto is not a single chain. But he underestimates the engineering – and trust – required to make that future work. Education is the new mining rig for the mind. We need to teach the next generation of developers and infrastructure builders to think in terms of orchestration, not just performance. The winning team won’t be the one that builds the fastest chip or the most scalable chain. It will be the one that builds the simplest, most reliable orchestration layer for a diverse set of trust and compute resources. As the market sleeps, architects of this new interoperability layer wake up. I’ve been there building BlockJakarta, training 200 developers in smart contract auditing and compliance. The same patterns emerge: you can’t trust a single point of failure. You need a combination of solutions – and you need to test them relentlessly.
So here is my forward-looking judgment: In five years, the inference market will look like the L2 ecosystem today – many providers, each claiming to be the fastest for a specific use case, all interconnected by a thin orchestration layer. The ISPs will become the equivalents of rollup-as-a-service platforms. Meanwhile, the question of “which chip is best” will be answered with “it depends” – the hallmark of a mature, modular industry. And those of us who lived through the crypto modular revolution will nod and say, “We told you so.” From core dev trenches to community heartbeat, this is how revolutions actually happen – not with a single grand theory, but with a thousand small, specialized solutions stitched together by relentless builders.
Art is the interface; blockchain is the canvas. And the inference chip is just a brush. The masterpiece comes from the combination.